About dashboard and statistics
Two read-only views sit beside Mission Control:
- The dashboard shows the repository's planned work: projects, tranches, backlog and tasks. Each task's status is derived on demand from forge facts. See the Overview.
- Statistics are comparable measures computed from Log events and forge facts: completion, cycle time, retries, revisions, interventions, gate failures, findings, escaped defects, usage, cost and reliability. Conclusions keep links to the observations they come from.
Statistics are descriptive. An answer reports what the Log recorded, never that one choice caused another.
How it works
The dashboard
The dashboard reads the forge (milestones, Issues and pull requests) and one file on disk, ..
Its pages go from the top level down to a tranche and a task:
| Page | Shows |
|---|---|
/studio | The top level |
/studio/tranches | Tranches |
/studio/backlog | The backlog |
/ | One project |
Statistics: reading the Log back
- 1
Read the source
The source is a log server (its read, stats and rejected routes, using
logs.readHeaders) or the log folder (<folder>/).<owner>- <repo>/ <work>. ndjson - 2
Normalise each line
Each stored line becomes exactly one dataset row or one quarantine record.
- 3
Cache the dataset
The dataset is cached at
<runtimeDir>/, never inside the repository.logs- cache/ cache. sqlite - 4
Answer the questions
Each question is a pure function over the dataset.
You fill, resume or rebuild the cache with vinaya sync. It never writes to the destination.
vinaya syncA sync run reads at most 50 pages of 1000 lines, and re-reads a 1000-position look-back by default. This keeps it inside a server's daily read allowance. See Telemetry.
The ten questions, plus a confidence comparison
| # | Question |
|---|---|
| 1 | Does a cheaper model finish? |
| 2 | Are reviewers strict? |
| 3 | What does a unit of work cost? |
| 4 | Where does the time go? |
| 5 | Does the product catch anything? |
| 6 | Which check catches most? |
| 7 | Are checks deterministic? |
| 8 | Which findings recur? |
| 9 | Is the instruction version the problem? |
| 10 | What change size converges? |
| — | The confidence comparison: does stated confidence predict review outcome? |
How answers stay honest
- A unit of work is the work reference (
workRef). A row without one is left out of every unit figure, and counted. - Every answer states its coverage:
rowsRead,lowTrustLeftOut,unitUnknown,rowsUsed,gaps,quarantinedandunknowns. - Unknown is never zero. A figure the Log cannot state is unknown, with its reason.
- Human labels are the only way in for judgement.
false_rejection,reversalandincidentare how a judgement the Log cannot capture enters an answer. Without them, the figures that depend on them read unknown ('no labels recorded'). - Gaps are never guessed. A span the source cannot fill is a gap. A run that saw a gap reports no deletion.
- No authority. Statistics read; they never feed control state.
Configuration
| Key | Effect |
|---|---|
logs.url, logs.readHeaders | The server source and its read credential, separate from the ingest logs.headers |
logs.folder | The folder source |
runtimeDir | Where the cache lives |
What gets logged
Nothing. The dashboard and statistics only read; vinaya sync is read-only on the destination.